Reading time: 12 minutes
When ChatGPT or another AI makes up information about your company, the error may have a source you can find. It may be a cited page, a directory, a review, an old site, a company with a similar name or old training data. So, to remove wrong information from ChatGPT, find the source and fix it there.
In short
- ChatGPT gets companies wrong in two ways: it repeats wrong data published on the web, or it fills a gap with a guess.
- The answer hints at the source: with a cited source, start with that page; with no source, the error comes from the model’s memory or a guess.
- In the OpenAI help pages we read, there is no channel for companies to correct answers. On Google, however, a company can edit its Business Profile and may claim its knowledge panel.
- After the fix, ask the same questions every month, because the error can come back when a source republishes the old data.
Why does ChatGPT make up information about companies?
ChatGPT makes up information about companies in two ways. In the first, it repeats wrong data that someone published, such as an outdated directory or an old page. In the second, it fills a gap with a plausible guess. After all, the model learned to predict words, not to check facts.
The error can even be a product feature that does not exist. In July 2025, Adrian Holovaty, founder of Soundslice, an interactive sheet music platform, shared a case like this. Screenshots of ChatGPT with ASCII tablature started to show up in the error logs. ASCII tab is a simple way to write music for guitar. ChatGPT was telling people to import this format into Soundslice, but the feature never existed.
“We’ve never supported ASCII tab; ChatGPT was outright lying to people.”
Adrian Holovaty, Soundslice (2025)
Because new users kept arriving with the wrong promise, the company decided to build the importer. When you can’t do the same, the way out is to find the source of the error. If the AI is making up information with no source, make the official version clear and easy to find.
The guessing mechanism, with the information AI gets wrong most often, is in the guide on why ChatGPT makes so many mistakes. Meanwhile, the types of errors and real cases involving companies are in the guide on what AI hallucination is.
Where does the wrong information in ChatGPT come from?
Before you fix anything, find out which path caused the error. The table sums up the possible sources, the sign each one leaves in the answer and where to fix it.
| Source of the error | Sign in the answer | Where to fix it |
|---|---|---|
| Page with the wrong data | The cited source has the error | With whoever published it |
| Directory or profile | Old address, hours or phone number | In the profile itself |
| Customer review | A complaint becomes a fact | In your reply to the review |
| Old website | A price, plan or service that changed | On your website |
| Company with a similar name | Data from another company | In the details that tell the two apart |
| Model training | No source, and the data used to be true | In current sources |
| Guess | No source, and the data never existed | On a clear official page |
How can you tell which source ChatGPT used for the wrong information?
When ChatGPT makes up information about your company, the answer itself shows whether the AI read a page or guessed. Ask again in a new chat, with search on, and open the cited sources. If one of them has the wrong data, that page is the source. An answer with no source, however, points to the model’s memory or a guess.
After all, OpenAI says search results and citations can be incomplete, outdated or incorrect. So OpenAI advises opening each cited source to check that it supports the answer. Without search, on the other hand, ChatGPT answers with what it learned in training, which has a cutoff date.
Then follow this order:
- The cited source has the wrong data. The page is the source: a directory, a news story, a review or your old site.
- The cited source does not have the data. The AI misread it or mixed up different pages. In this case, make your official page clearer and report the answer.
- There is no source, and the data used to be true. Wrong information from ChatGPT tends to come from training, such as a price from before the model’s cutoff date.
- There is no source, and the data was never true. Ask again with your city and CNPJ, Brazil’s company tax ID. If the error goes away, that signals a mix-up with another company; if it stays, the AI is probably making up information by guessing.
What if ChatGPT says your company has closed?
Start with the profiles that say whether your company is open. On a Google Business Profile, a location can show as temporarily or permanently closed, according to Google’s help. This status can also change through user suggestions.
According to Google’s page on profile updates, Google notifies the owner about some suggested edits. The owner then has 4 days to accept or reject them. If there is no reply, Google may publish the change when other public information supports it. In addition, Google Maps help says to mark a place as permanently closed if the business has changed its name significantly. In other words, a rebrand can turn into a closure.
Outside Google, look for news about a location closing. In Brazil, also check CNPJ lookup sites for a branch listed as closed. Check, too, for companies with a similar name that have closed. Finally, reopen the profile, reject wrong suggestions and state on your website, with address and hours, that the company is open.
Where does ChatGPT get outdated price information?
First, look for where the old price is still published. Search for the price in quotes on Google and Bing, then repeat the search with site: and your domain. Old pricing pages, PDFs, launch posts, comparison sites and marketplaces may hold the wrong information that ChatGPT repeats.
If the price was once on your site, the Wayback Machine, from the Internet Archive, may show archived versions of the page. If no page publishes the price and the answer cites no source, the data tends to come from training. In that case, make the current price clear, with the date, on your website and profiles. The steps to clean up old pages are in the guide on how to fix an old address, price or service.
Can you ask OpenAI or Google for a correction?
When ChatGPT makes up information about your company, the OpenAI help pages we read show no direct way to ask for a fix. The thumbs down sends feedback that may go into training, but those pages do not promise to fix the answer. Google, on the other hand, offers more paths, such as the Business Profile and the knowledge panel.
According to OpenAI, a thumbs up or thumbs down sends feedback. On individual plans, OpenAI may use that conversation to train its models. However, the pages we read give no deadline and no channel for companies to correct wrong information in ChatGPT. The formal correction request covers personal data, such as wrong data about a business partner. In the documentation, we also found no company profile to claim in ChatGPT. The steps to report an error are in the guide on how to report a ChatGPT error.
How do you fix your Google knowledge panel?
The knowledge panel is the summary box Google shows for entities in its Knowledge Graph, such as companies and people. According to Google, it generates these panels automatically, with information from many sources across the web. AI Mode can also tap into the Knowledge Graph, according to the feature’s 2025 announcement.
To fix it, someone who represents the company clicks Claim this knowledge panel. Then they sign in to an official profile linked to it, such as Search Console or the YouTube channel. Next, under Suggest edits, they describe what is wrong and add public links that support the right data. According to the knowledge panel help, Google reviews feedback from verified users within a few days, sometimes longer.
Some panels don’t have the claim option yet. In that case, Google’s help says to click Feedback and send the suggested fix. If the panel mixes your company with another one with a similar name, the advice is to contact Google support. For a business with a physical location, however, the path is the Business Profile.
When do the courts step in?
When AI makes up serious information, such as a link to a scam, legal action is possible. However, it depends on each country’s laws. In Germany, a Munich court granted two publishers a preliminary injunction against the operator of a search engine. The court’s press release does not name the company, but the German news agency epd says it is Google. According to epd, the decision came in May 2026 and bars spreading false claims about the publishers in AI Overviews. The agency also says the AI mixed data on dubious companies with the publishers’ data and tied them to scams. The decision is not final. In Brazil, review the case with your legal team before you act.
How do you fix it at the source: website, profiles, directories and Wikipedia?
To get wrong information out of ChatGPT, first fix what your company controls, such as your website and profiles. Then ask whoever published the wrong data to correct it, such as directories, media outlets and review sites. Use the same data everywhere, with the date of the last review. That way, the AI finds official information to compare.
- Website: fix every page with the old data, including PDFs and posts. Then redirect what you took down and request a new crawl.
- Google Business Profile: edit your verified profile and review the changes Google made, which show in blue in the editor. When a suggested edit arrives, reply within 4 days.
- Directories, marketplaces and CNPJ lookup sites: fix the ones where your company has an account and ask the others to correct the data, with a link to your official page. For address, status and partners, first check that your CNPJ record with Brazil’s Federal Revenue Service (Receita Federal) is up to date.
- Company with a similar name: repeat your legal name, CNPJ, city and what the company does on your website and profiles. Then test the question with and without these details.
If ChatGPT or another AI makes up information with no source and no page has the error, the fix is on your website. That is why the official page needs to answer the question with the right data and the date.
What if the error is in a review, the press or Wikipedia?
An old complaint, such as “they don’t deliver on Saturdays,” can become a fact in the AI’s answer. Reply to the review in public, with the right data and without arguing. On Google, an approved reply shows below the review, and the customer gets a notification. The company can also flag reviews that violate Google’s content policies, according to Business Profile help.
For the press, send the outlet proof of the right data and ask for a correction. On Wikipedia and Reclame Aqui, a Brazilian complaint site, follow each one’s rules, summed up in the guide on the most cited sources in ChatGPT.
How long does it take for wrong information to leave ChatGPT?
There is no single timeline, because the time depends on where the error came from. When it comes from a page, the answer with search can change after the search engine reads the fixed version. When it comes from training, however, the error can last until a newer model goes into use.
| Where the error was | What sets the timeline |
|---|---|
| Your page or a third-party page | A new crawl by the search engine |
| Google Business Profile | Google’s review |
| Knowledge panel | The review of your feedback |
| Model training | A newer model in use |
According to Google, a new crawl can take from a few days to a few weeks. For Bing, which Microsoft says powers ChatGPT, IndexNow reports the change right away, though it does not guarantee indexing. If ChatGPT is making up information from training, however, the AI may repeat the old data until the model changes.
How do you track whether ChatGPT is making up information again?
To know if wrong information is back in ChatGPT, keep a log of each error, with its source and fix. Then ask the same questions every month in the same AI engines. The error can return when a source republishes the old data or another page copies the wrong version. A new model can also bring the error back.
- Log the date, engine, question, answer, cited source, origin, fix made and status.
- Run each question several times, with and without search, because the answer can change with each run.
- Create an alert for the wrong data, in quotes. According to Google, Google Alerts sends emails when new results on a topic show up.
- Test again when a new model arrives in ChatGPT, since the training changes.
- Go back to the source whenever the answer cites a page with the old data.
If ChatGPT makes up information again, the AI may be reading a new source, so compare it with the first one. The data collection protocol, with runs and margin of error, is in the guide on how to measure what AI says about your brand.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- OpenAI, Searching the web with ChatGPT (2026).
- OpenAI, How your data is used to improve model performance (2026).
- Google, Understand Google updates on your Business Profile (2026).
- Google, Mark your business as closed (2026).
- Google, Submit feedback on content about you, in the knowledge panel help (2026).
- Adrian Holovaty, Adding a feature because ChatGPT incorrectly thinks it exists (2025).
- Landgericht München I, press release on liability for AI Overviews (2026).
- epd, Landgericht München: Google muss für KI-Übersicht haften (2026).
